Real World Action: Inter-Body Artificial Intelligence
Executive Summary
II. Real World Action: Inter-Body Artificial Intelligence
A. Inter-Body Artificial Intelligence
Inter-body intelligence asks what the estimated body configuration means physically and tactically. Pose recognition locates an arm; relational interpretation determines whether that arm is posting on the mat, framing against a torso, maintaining a grip, trapped beneath a body, or being isolated for a submission. These relationships determine which actions are available and which constraints an action must respect.
One possible implementation is a typed relational graph. Nodes represent relevant body regions and environmental supports such as the mat. Edges represent relationships such as contact, grip, hook, frame, support, loading, pinning, or entanglement. Each relationship can carry direction, confidence, persistence, relative motion, and available force or pressure estimates. A contact edge records that two regions touch; a control relationship requires additional evidence about how that contact limits movement.
Relationships are inferred from multiple signals over time. A hand touching the mat is a candidate post; evidence that it bears load strengthens that interpretation. A forearm contacting a torso is a candidate frame; its orientation, mechanical resistance, and effect on relative motion help determine whether it functions as one. Visual contact alone should not be treated as proof of grip strength, load transfer, or positional control.
The relational state also represents the interaction between support, balance, and mobility. The system can estimate which regions bear weight, which limbs remain free to move, and how a proposed movement would change the participants’ support relationships. During grappling, support can pass through the other participant as well as directly to the mat, so evaluating each body in isolation can miss important constraints.
For example, two participants may remain in the same broadly recognized guard position while a wrist grip is released, a foot becomes a hip frame, or a supporting hand lifts from the mat. The position label may remain unchanged, but the relational state—and therefore the available next actions—has changed. The robot can use those changes to reconsider whether to continue, modify, abandon, or transition from its current technique.
The resulting control loop is observation, body-state estimation, relationship inference, candidate-action evaluation, and reassessment after movement. Candidate actions are evaluated against the current relationships, sport rules, training objectives, and execution constraints. Tap detection and other termination conditions must be able to preempt that loop. These are proposed technical implementation paths for the architecture, rather than a statement that a complete autonomous grappling system has already been demonstrated.
B. Techniques Become Transformations
In this architecture, a Brazilian jiu-jitsu technique can be represented as a transformation between relational states.
An armbar is not merely a video or stored joint trajectory. It is a progression from an entry condition through intermediate control relationships toward a target state. A guard pass changes the relationships among the robot’s limbs, the participant’s limbs, the mat, the participants’ centers of mass, and their respective bases of support.
The robot can evaluate candidate actions according to whether they improve position, create leverage, isolate a body region, remove a supporting relationship, reduce mobility, disrupt balance, or establish the entry condition for a subsequent technique.
When the participant reacts, the system observes the resulting relationships and decides what comes next.
That is the beginning of genuine physical interaction intelligence.
C. High-level armbar states
The starting node is DEEP ARMBAR ATTEMPT. The diagram describes selected outcomes from the attacking robot’s perspective. GUARD PASSED means the opponent passes the robot’s guard.
D. What does a 0.60 probability mean?
The estimate is specific to an observed state and a proposed continuation policy. A deep armbar attempt can indicate established control and reduced defensive mobility. Evidence may include knee height relative to the opponent’s shoulders, leg arrangement, hip position, elbow location, persistent grips, and measured contact. Interlocked feet or high knees alone do not establish effective control.
The model may also consider the robot’s previous outcomes from comparable configurations, this opponent’s observed defenses, changes in posture or support, and confidence in hidden-limb estimates.
The illustrative 0.60 means the estimated likelihood of reaching TAP / RELEASE from this particular state under the selected continuation policy. It is not a universal armbar success rate and does not mean the opponent is 60% compromised. Physical constraint is an input; the probability describes a predicted outcome.
The 0.40 branch represents the modeled alternative. After ARMBAR DEFENDED, 0.35 and 0.65 are conditional estimates for the next outcomes from that later state. This simplified diagram omits ongoing exchanges and independent stop conditions. Deployment requires explicit outcome definitions, an evaluation horizon, handling of unresolved and stopped exchanges, and calibration against measured results.
E. Detailed branch diagrams
The planned completion diagram expands DEEP ARMBAR ATTEMPT → TAP / RELEASE into control maintenance, reassessment, submission detection, and release confirmation.
The planned defense diagram expands DEEP ARMBAR ATTEMPT → ARMBAR DEFENDED into changes in posture or control, defensive progress, and reassessment of available transitions. These branch diagrams remain under development; internal probabilities need their own context.
F. From state estimation to action
FUSED POSE AND PHYSICAL STATE → INTER-BODY RELATIONAL STATE → BJJ LLM → BJJ ACTION TOKEN TIMELINE → VALIDATION AND MOTION CONTROL
The robot identifies its current state, evaluates candidate actions and predicted outcomes, then selects an intended transition. The opponent’s response and physical exchange determine the actual next state. A continuation policy cannot command the opponent to tap.
The BJJ LLM: From Relational State to Action Token
Section II uses the BJJ large language model (LLM) trained and adapted in section III. Its runtime inputs include the fused pose and physical-state estimate, inter-body relationships, recent interaction history, candidate transitions, training objectives, and applicable execution constraints. A structured state representation supplies the model with grounded observations and their confidence.
The BJJ LLM selects a proposed response consistent with its learned instructor persona and outputs a structured BJJ action token, such as MAINTAIN ARM CONTROL, ADJUST RELATIVE POSITION, or RELEASE. The output identifies the intended control relationship, relevant body regions, timing, limits, and conditions requiring reassessment. Success estimates remain tied to the observed state and proposed continuation, with calibration required before they are treated as measured probabilities.
Section III supplies the learned methodology; section II applies it to the current exchange. Downstream validation and motion control determine whether and how a proposed token can execute. Updated sensing feeds back into the model’s next decision, while tap and termination handling can interrupt the timeline independently.
G. What is a BJJ action token?
A BJJ action token is a structured, machine-readable output of the BJJ LLM representing an executable part of its selected response. It specifies a control objective, relevant body regions, timing, constraints, and reassessment or termination conditions. A token may be encoded using multiple model tokens; it is a semantic action unit rather than necessarily a single language-model token.
Example token: MAINTAIN ARM CONTROL
Controlled regions: robot hands and estimated opponent forearm. Objective: maintain the existing grasp and relative alignment. Constraints: applicable force, speed, and joint limits. Timing: start condition and maximum duration. Reassessment triggers: grip slip, opponent movement, or declining confidence. Termination conditions: tap, verbal submission, or another stop signal.
The proposed execution layer checks the BJJ action token against current physical state and execution constraints before a downstream motion controller converts its objective into coordinated joint commands. Feedback can invalidate or replace a pending token, and termination signals preempt execution. The LLM proposes the action objective; feedback control executes it.
An illustrative timeline is MAINTAIN CONTACT → ADJUST RELATIVE POSITION → REASSESS CONTROL → CONTINUE, TRANSITION, OR RELEASE. Tokens may overlap where appropriate, such as maintaining contact while adjusting position. Fresh observations can interrupt, replace, or cancel pending tokens.
H. Actuator Design and Physical Execution for BJJ
BJJ physical execution requires controlled resistance, yielding, sustained contact, and coordinated movement under changing external loads. Proposed engineering priorities include backdrivability, low limb inertia, torque feedback, usable joint range, and reliable release. These are design directions requiring platform-specific validation.
Brushless, planetary, and linear describe different parts of an actuator. Brushless describes the motor; planetary describes a gear transmission; linear describes translational output. A brushless motor can drive a planetary gearbox or a screw-based linear actuator. Selection therefore concerns the complete actuator and transmission, not a choice among three mutually exclusive technologies.
A proposed starting point for major joints is a brushless motor with low-ratio reduction and torque control, often described as quasi-direct drive. A low-ratio planetary transmission is one candidate. The objective is responsive interaction and the ability to yield when externally moved. Gear ratio, friction, rotor inertia, backlash, packaging, and continuous torque must be evaluated together; planetary gearing alone does not guarantee backdrivability.
Series-elastic actuation adds a compliant element between the drive and load. Selected joints could use this arrangement to measure transmitted torque through spring deflection and accommodate contact transients. Compliance introduces deflection and control-bandwidth tradeoffs, so stiffness and damping should be selected for each joint and task.
Linear actuators may suit selected knee, hip, or torso linkages when packaging and force requirements justify them. Their suitability depends on stroke, linkage leverage across the full motion, speed, friction, and ability to release under load. Self-locking drives are poor default choices where a limb must yield promptly. Rotary actuation is a practical baseline for joints requiring broad angular motion.
Mechanical adaptations should address hip rotation and flexion for guard work, shoulder mobility for frames and posts, torso articulation for turning and hip escapes, and wrist and finger compliance for grips. Joint limits must be assessed under load. Structural bearings and load paths should withstand externally applied twisting and side loads, rather than relying on motor shafts or gear teeth to carry every grappling load.
Light distal limbs reduce rotational inertia. Moving heavier drives closer to the torso may help, although tendon or cable transmissions add friction, stretch, routing, and maintenance tradeoffs. Compliant contact surfaces, protected pinch points, and releasable fingers support controlled interaction. Sustained holds require continuous-torque and thermal assessment, including cooling beneath padding and a gi; peak torque alone is insufficient.
The BJJ LLM trained in section III proposes action tokens in section II. An execution layer checks each token against the current physical state and hardware capabilities, then a feedback controller coordinates joint torque, motion, stiffness, and damping. For MAINTAIN ARM CONTROL, the controller can maintain bounded contact while allowing compliant movement and reassessing grip slip. The LLM does not directly close the motor-control loop.
The available physical output changes with joint posture, temperature, contact, loading, and remaining actuator capacity. Those limits should feed back into action selection so the model can choose a feasible continuation or transition. Tap and termination handling must preempt pending actions and coordinate release without an uncontrolled collapse. Evaluation targets include release latency under load, backdrive resistance, torque-tracking accuracy, contact-force overshoot, sustained-hold temperature, and reachable BJJ configurations.
Related architecture details
Proposed Architecture
Technical Specifications
Partner with
TrustRobotics
Integrate the industry's premier
AI-to-actuator safety layer.
Accelerate Open Development
Validate & Test Stack
Secure Commercial Licensing